Routing hypergraph convolutional recurrent network for network traffic prediction

Effectively predicting network traffic is a fundamental but intractable task in IP network management and operations. Many methods that can capture complex spatiotemporal dependencies from network topology and traffic sequence data have achieved remarkable results and become dominant in this task. However, the previous methods seldom consider the spatial information from the routing scheme, which also determines the flow direction and trend of network traffic. To fill this gap, we regard a routing path as a hyperedge and utilize a hypergraph instead of a simple graph to model network node connections based on the routing relevance. Then, we propose a novel multi-step network traffic prediction model named routing hypergraph convolutional recurrent network (RHCRN), which is built on the seq2seq structure with the hypergraph convolutional recurrent unit (HCRU). The HCRU is composed of the 2-layer hypergraph convolutional network (HGCN) and gated recurrent unit (GRU). The node-edge-node transform process of the HGCN layer is ideal for exploring the complex spatial correlation between the routing paths and network nodes. The GRU is used to extract the temporal correlation from dynamic network traffic data. Extensive experiments on three real-world IP network datasets demonstrate that our model is robust and outperforms other advanced baseline models.